CN105760365A - Probability latent parameter estimation model of image semantic data based on Bayesian algorithm - Google Patents
Probability latent parameter estimation model of image semantic data based on Bayesian algorithm Download PDFInfo
- Publication number
- CN105760365A CN105760365A CN201610142356.8A CN201610142356A CN105760365A CN 105760365 A CN105760365 A CN 105760365A CN 201610142356 A CN201610142356 A CN 201610142356A CN 105760365 A CN105760365 A CN 105760365A
- Authority
- CN
- China
- Prior art keywords
- probability
- document
- semantic
- algorithm
- image
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
- G06F16/5846—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content using extracted text
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
- G06V30/26—Techniques for post-processing, e.g. correcting the recognition result
- G06V30/262—Techniques for post-processing, e.g. correcting the recognition result using context analysis, e.g. lexical, syntactic or semantic context
- G06V30/274—Syntactic or semantic context, e.g. balancing
Abstract
The invention provides a Bayesian probability estimation model, aiming at the problem of a semantic gap in image search at present. Probability latent semantic subject term (hereinafter referred to as a subject term) parameters are added on the basis of image semantic retrieval to establish a document-subject term-semantic characteristic word probability relation; a Bayesian probability estimation algorithm theory is combined to calculate posterior probability, wherein the posterior probability is equal to an arithmetic product of prior probability and a likelihood function; a maximum likelihood estimated value of the posterior probability is solved by using an EM algorithm, so that the Bayesian probability estimation model is established. In view of the document-subject term-semantic characteristic word relation, a document is mapped, and the document corresponds to an image, so that the image needed by a user is finally displayed.
Description
Technical field
The present invention relates to about Bayesian probability algorithm for estimating.
Background technology
Theory of probability is the rigorous inferential strong mathematics branch of logic, and Bayesian formula is the most important in theory of probability
Formula, many terms that Bayes is used are used so far.The Major Difficulties of machine learning be " being set forth " morphology and
The semantic difference " really expressed ".Produce the reason of this problem mainly: 1. a word can have multiple meaning with many
Individual usage.2. synonym and near synonym, and according to different linguistic context or other factors, different words is also possible to represent phase
The same meaning.
The Major Difficulties utilizing Bayesian probability estimation principle to carry out computing machine study is very easily, is also current optimum
Elegant method.
The content characteristic retrieval of image, is up to the present the most ripe a kind of searching system, but his retrieval effectiveness is
Unsatisfactory.Subject matter is exactly problem produced by machine learning, and the present invention utilizes Bayesian probability algorithm for estimating exactly
The problem solving machine learning.
Summary of the invention
The present invention carries out semantic number by Bayesian probability latent semantic analysis to the semantic feature of unsupervised learning image
According to excavation, set up the model of a probability potential applications algorithm.Mainly solve the subject matter 1 in image retrieval, semanteme
Wide gap；2, adopted many words and polysemy produced retrieval data are not in place and repeated retrieval.
If the corresponding document of piece image, if the key words of image, semantic feature correspondence document, set up a document-
Key words semantic vector co-occurrence matrix.According to Bayesian probability estimation principle, the computational methods of conditional probability, applied probability is dived
The maximum similarity of document is calculated in descriptor.
The technical scheme is that Bayesian Estimation algorithm based on machine learning sets up probability latent semantic model, bag
Include following steps:
Step 1, is excavated without supervision image, semantic knowledge base data by probability latent semantic analysis method (PLSA), builds
Probability semantic relation between vertical document (image)---potential applications---word (image, semantic feature), it is simply that each group (D, W)
All relevant with potential applications Z；
Step 2, calculates potential applications main body word maximal possibility estimation and expectation maximization with Bayesian probability algorithm for estimating；
Step 3, sets up probability latent semantic model.
And, in step 1, unsupervised learning is a kind of based on image, semantic knowledge base, and its generation is to be used by computer
The linguistic indexing of pictures method of gauss hybrid models (Gaussian Mixture Model, GMM) automatically generates；
And, in step 2, when Bayesian Estimation calculates, prior probability distribution meets BETA distribution, according to prior probability, seemingly
So function calculates posterior probability；
And, in step 2, potential function Z maximal possibility estimation is trained, and algorithm the more commonly used in maximal possibility estimation is just
It it is expectation-maximization algorithm.Expectation-maximization algorithm is divided into two steps:
1. the estimation of Expectation Step implicit parameter
2. Maximization Step determines actual parameter, then does maximal possibility estimation according to actual parameter.
Accompanying drawing explanation
Accompanying drawing 1 potential descriptor figure.
Accompanying drawing 2 asymmetric probability latent semantic model.
The symmetrical probability latent semantic model of accompanying drawing 3.
Accompanying drawing 4 implements schematic diagram for the present invention.
Detailed description of the invention
Technical solution of the present invention is described in detail below in conjunction with drawings and Examples.
The present invention proposes to supervise nothing based on Bayesian probability algorithm the parameter of the probability potential applications of image, semantic data
Estimate model, it is assumed that image, semantic knowledge base by computer with gauss hybrid models (Gaussian Mixture Model,
GMM) linguistic indexing of pictures method automatically generates, the most independent between mark semanteme, tandem useless.In order to the most detailed
Thin elaboration specific embodiments of the present invention, are further described below based on accompanying drawing, and implementing procedure is as follows.
Step 1, if the corresponding document space of piece image, the keyword space of the Feature Mapping of image to document.Make
A document, i.e. one vector being made up of word is represented with bag-of-word model.Word, independent of text semantic, is neglected
Omit word order in a document, the frequency that only performance word occurs in a document, a dimension of a word representation space
Degree, thus obtain document word co-occurrence matrix, each item of this matrix represents i-th word in jth piece document
Weight.Co-occurrence matrix has passed through term frequency-inverse document frequency (TF IDF) dimensionality reduction.
Step 2, sets up the probabilistic relation between document, document key word and potential scenic themes word, Fig. 1.By generally
Rate latent semantic analysis sets up PLSA probabilistic model figure, Fig. 2.Wherein observed data D represents document, and W represents one in document
Feature word, potential applications descriptor Z is non-observed data.D=(d1, d2,..., di) there is the document sets of N piece document；Z=(Z1,
Z2,..., Zk) document have K hidden variable descriptor, theme word set；W=(W1, W2,..., Wj) document have M
Word, set of words.Generating a document and the text set of word, co-occurrence matrix A=[P (di, Wj)] | D | × | W | represents, d generation
Table a certain piece document, W represents some word, and (di, Wj) represents document di and word the Wj frequency of occurrences, namely P simultaneously
(di, Wj) is the number that two variablees occur simultaneously.Every a line of matrix A represents a document, and every string represents a list
Word.Then P (di, Wj) is an observable number, according to the mapping of Fig. 1, introduces unobservable hidden variable Z, and Z makes
Document and word are with good conditionsi separate, and Z belongs in hidden variable collection Z, and wherein the value of K is sized depending on experience, on the one hand
Wish that K is sufficiently large, it is possible to be suitable for all of potential applications structure, if but excessive, it is easily introduced noise, will produce for using
Certain impact；If too small, then can not manifest error and other minor details of sample, generally take 20 to 100
Between.A (di) from N(i=1,2,3 ... ..N) piece document sets selects the probability of a document di；A (di, Wj) is that word Wj exists
The frequency occurred in document di；A (Wj | Zk) represent under conditions of unobservable hidden variable Zk determines, word wj goes out
Existing conditional probability；A(Zk| di) represent that document di belongs to kth under conditions of the probability of occurrence of given i-th document
The probability of descriptor；
Utilize Bayesian probability algorithm for estimating, can the generation model of the co-occurrence data of defined terms-document according to the following procedure:
1, first from document sets, select a document di with probability A (di)；
2, again with probability A (Z in document dik| di) select a hiding theme class variable ZkProbability；
3, determine hiding class variable ZkAfter, with probability A (Wj | Zk) select a word Wj probability.So can be obtained by one
Individual observed value (di, Wj), and hidden variable Z not being observedkIt is dropped.Assume that document d and word w are at given theme list
It is satisfied with independent same distribution, say, that implicit class Z and word W are not dependent on what specific document D generated under conditions of word z.
Then data above generation process can be stated by following Bayesian probability estimation formulas:
A (di, Wj)=A (di) A (Wj | di) (1)
Wherein: A (di): be prior distribution, A (Wj | di): it is likelihood function, A (di, Wj): be Posterior distrbutionp.Bayes side
Method is thought, the data of learning parameter=priori+observe.The prior distribution of priori=parameter choose+distributed constant
Choose.According to bayesian theory, in order to reach the distribution that Posterior distrbutionp and prior distribution belong to same, choosing prior distribution is
Beta is conjugated distribution:
(2)
2 formulas are substituted into 1 formula obtain
(3)
Observe Fig. 2 to describe and (2 formula), according to the feature of conjugation distribution, so they are symmetric function, Fig. 2 model conversation be
Model shown in Fig. 3.By the general A of bayes rule turn around condition (z/d), (1) can be rewritten as equation below
(4)
According to (1) formula, A (di, Wj) is that posteriority 2 is distributed, A (Z1) it is prior distribution, and A (Wj | Zk) A(di| Zk) be
Likelihood function.Prior distribution Z1Being uniformly distributed, according to Bayesian assumption, A (Z on interval (0,1)1)=1.So from (4) formula
In we to obtain sample (di, Wj) be the linear combination about implicit variable Z, be made up of two multinomial distribution, two
The parameter of multinomial distribution is A (W | Z) and A (d | Z), and namely probability latent semantic model is former according to Maximum-likelihood estimation
Reason, by asking for the maximum of following log-likelihood function, obtains the model parameter of potential applications.
Step 3, existing maximal possibility estimation comes computation model parameter A (W | Z) and A (Z | d), the public affairs of likelihood function
Formula:
(5)
Here the word frequency of all words that (di) occurred in being document di and, his value is to model parameter A (W | Z) and A (d
| Z) there is no any impact.Upper examination can be write as
(6)
Owing to symbol of suing for peace in this object function occurs in integration, therefore this letter
Number is difficult to maximize, in order to solve this problem, it is possible to use EM algorithm.
Step 4, (Expectation Maximization Algorithm, translates again expectation maximization to EM algorithm
Algorithm), it is a kind of iterative algorithm, for the maximum likelihood of the probability parameter model containing hidden variable (hidden variable)
Estimate or maximum a posteriori estimate.In statistical computation, greatest hope (EM) algorithm is at probability (probabilistic) mould
Finding parameter maximal possibility estimation or the algorithm of MAP estimation in type, wherein probabilistic model depends on cannot observe hidden
Hide variable (Latent Variable).EM algorithm alternately calculates through two steps:
The algorithm flow of EM is as follows:
1. initialize distributed constant
V=A (Wj | Zk) } m*k, U={A (di | Zk) } n*k
With square error criterion K-Means algorithm, calculate the initial value of U, V parameter respectively.Idiographic flow:
(1), select N number of sample to be divided into disjoint K subset, calculate average α of each subset1,α2,......,αkAnd β,
If i-th subset has Ni sample, totalIf Si represents i-th subset.Wherein:
β is error sum of squares clustering criteria, and it is k cluster centre α1,α2,......,αkRepresent k sample set S1,
S2 ... produced total square-error during Sk, the cluster making β minimum is exactly the optimum knot under error sum of squares criterion
Really；
(2), an alternative sample x is randomly choosed, if this；
(3) if Ni=1, then forward to (2), the most discontinuously；
(4), calculate；
(5), for j=1 used, 2 ..., if there is certain in kMake, then x is moved on to from Si
In St；
(6), α is recalculatediValue, and revise β；
(7), β value keep constant, then iteration terminates, and otherwise continues iteration, turn (2).The result calculated is put into U and V respectively
Middle as initial value；
2. repeat until restraining:
The first step is to calculate expectation (E), utilizes the existing estimated value to hidden variable, calculates its maximum likelihood estimator；
(7)
The formula left side represents under the Probability Condition i-th document, jth word occur, obtains existing kth and implies the general of theme
Rate, it is understood that for containing document and the jth word contribution degree to kth theme of k theme probability given one,
I-th document obtaining existing jth word belongs to the probability of kth theme.Denominator on the right of formula represents to be needed to join model
Two products of number A (W | Z) and A (Z | d) are normalized that (concrete normalized mode is: belong to the document
Arbitrarily theme and the product of this theme contribution degree is rescued by this word equally and be molecule, then belongs to specific master with document
Probability and the word of topic are denominator to this particular topic contribution degree, ask business.Second step is to maximize (M), maximizes
The maximum likelihood value tried to achieve in E step is to calculate the value of parameter:
(8)
Calculate and give the word j contribution degree to implicit theme k:
(9)
Calculate and give the probability that document (i-th document) belongs to implicit theme k:
(10)
K=K in implicit theme, when being recycled toTime stop calculate.(5th) formula reaches maximum.Finally obtain
Degree of association between the joint probability of (di, wj), i.e. (di, wj).
Claims (1)
1. the technical scheme is that based on Bayesian probability algorithm the probability potential applications without supervision image, semantic data
Parameter estimation model, its feature is, use Bayesian Learning Theory basic concept for excavate without supervision image, semantic
Potential relation between data:
Step 1, sets up image document---potential applications Feature Words---word, the probabilistic relation between image, semantic feature three；
Step 2, uses Bayesian probability algorithm for estimating, and posterior probability=prior probability * likelihood function, with Bayesian assumption to priori
Probability carries out paying value；
Step 3, utilizes EM algorithm to seek the maximum likelihood estimator of posterior probability；
Step 4, carries out EM initial value setting with K-m algorithm；
Step 5, E step calculates desired value；
Step 6, M step calculates parameter maximum likelihood estimator；
Step 7, sets loop parameter K, and k < continues EM to calculate, terminates to calculate during k=K during K.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610142356.8A CN105760365A (en) | 2016-03-14 | 2016-03-14 | Probability latent parameter estimation model of image semantic data based on Bayesian algorithm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610142356.8A CN105760365A (en) | 2016-03-14 | 2016-03-14 | Probability latent parameter estimation model of image semantic data based on Bayesian algorithm |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN105760365A true CN105760365A (en) | 2016-07-13 |
Family
ID=56333118
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610142356.8A Pending CN105760365A (en) | 2016-03-14 | 2016-03-14 | Probability latent parameter estimation model of image semantic data based on Bayesian algorithm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105760365A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108446273A (en) * | 2018-03-15 | 2018-08-24 | 哈工大机器人（合肥）国际创新研究院 | Kalman filtering term vector learning method based on Di's formula process |
| CN108932270A (en) * | 2017-05-27 | 2018-12-04 | 福建省农业科学院果树研究所 | Contrast method is retrieved based on the loquat category germ plasm resource of Bayes and feedback algorithm |
| CN109615608A (en) * | 2018-11-13 | 2019-04-12 | 昆明理工大学 | A kind of method of natural image Bayesian Reconstruction during human brain is movable |
| CN112487185A (en) * | 2020-11-27 | 2021-03-12 | 国家电网有限公司客户服务中心 | Data classification method in power customer field |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20030219146A1 (en) * | 2002-05-23 | 2003-11-27 | Jepson Allan D. | Visual motion analysis method for detecting arbitrary numbers of moving objects in image sequences |
| CN102902976A (en) * | 2011-07-29 | 2013-01-30 | 中国科学院电子学研究所 | Image scene classification method based on target and space relationship characteristics |
| CN103077530A (en) * | 2012-09-27 | 2013-05-01 | 北京工业大学 | Moving object detection method based on improved mixing gauss and image cutting |
| CN104036021A (en) * | 2014-06-26 | 2014-09-10 | 广西师范大学 | Method for semantically annotating images on basis of hybrid generative and discriminative learning models |
-
2016
- 2016-03-14 CN CN201610142356.8A patent/CN105760365A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20030219146A1 (en) * | 2002-05-23 | 2003-11-27 | Jepson Allan D. | Visual motion analysis method for detecting arbitrary numbers of moving objects in image sequences |
| CN102902976A (en) * | 2011-07-29 | 2013-01-30 | 中国科学院电子学研究所 | Image scene classification method based on target and space relationship characteristics |
| CN103077530A (en) * | 2012-09-27 | 2013-05-01 | 北京工业大学 | Moving object detection method based on improved mixing gauss and image cutting |
| CN104036021A (en) * | 2014-06-26 | 2014-09-10 | 广西师范大学 | Method for semantically annotating images on basis of hybrid generative and discriminative learning models |
Non-Patent Citations (4)
| Title |
|---|
| THOMAS HOFMANN: "Unsupervised Learning by Probabilistic Latent Semantic Analysis", 《MACHINE LEARNING》 * |
| 付杰: "基于语义的图像标注关键技术研究", 《中国科技论文在线》 * |
| 唐湘晋，陈家清，毛树华主编: "《应用数理统计》", 30 October 2013, 武汉：武汉工业大学出版社 * |
| 李志欣等: "基于概率主题建模的图像语义学习与检索", 《广西师范大学学报：自然科学版》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108932270A (en) * | 2017-05-27 | 2018-12-04 | 福建省农业科学院果树研究所 | Contrast method is retrieved based on the loquat category germ plasm resource of Bayes and feedback algorithm |
| CN108932270B (en) * | 2017-05-27 | 2021-11-30 | 福建省农业科学院果树研究所 | Loquat germplasm resource retrieval contrast method based on Bayes and feedback algorithm |
| CN108446273A (en) * | 2018-03-15 | 2018-08-24 | 哈工大机器人（合肥）国际创新研究院 | Kalman filtering term vector learning method based on Di's formula process |
| CN108446273B (en) * | 2018-03-15 | 2021-07-20 | 哈工大机器人（合肥）国际创新研究院 | Kalman filtering word vector learning method based on Dield process |
| CN109615608A (en) * | 2018-11-13 | 2019-04-12 | 昆明理工大学 | A kind of method of natural image Bayesian Reconstruction during human brain is movable |
| CN112487185A (en) * | 2020-11-27 | 2021-03-12 | 国家电网有限公司客户服务中心 | Data classification method in power customer field |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Balažević et al. | Tucker: Tensor factorization for knowledge graph completion | |
| Siddiqi et al. | Keyword and keyphrase extraction techniques: a literature review | |
| Ch et al. | Bayesian learning | |
| Daumé III | Bayesian multitask learning with latent hierarchies | |
| Gupta et al. | Beyond nouns: Exploiting prepositions and comparative adjectives for learning visual classifiers | |
| CN108595706A (en) | A kind of document semantic representation method, file classification method and device based on theme part of speech similitude | |
| US20070294241A1 (en) | Combining spectral and probabilistic clustering | |
| CN105760365A (en) | Probability latent parameter estimation model of image semantic data based on Bayesian algorithm | |
| Song et al. | Open domain short text conceptualization: A generative+ descriptive modeling approach | |
| Hammar et al. | Deep text mining of instagram data without strong supervision | |
| Inouye et al. | Admixture of Poisson MRFs: A topic model with word dependencies | |
| Melnykov et al. | Semi-supervised model-based clustering with positive and negative constraints | |
| Osborne et al. | Encoding prior knowledge with eigenword embeddings | |
| US11748567B2 (en) | Total correlation variational autoencoder strengthened with attentions for segmenting syntax and semantics | |
| US20220043975A1 (en) | Disentangle syntax and semantics in sentence representation with decomposable variational autoencoder | |
| Pighin et al. | On reverse feature engineering of syntactic tree kernels | |
| Muthukrishnan et al. | Simultaneous similarity learning and feature-weight learning for document clustering | |
| Lees et al. | Embedding semantic taxonomies | |
| Wu et al. | A modified fuzzy c-means algorithm for collaborative filtering | |
| Ono et al. | Person name disambiguation in web pages using social network, compound words and latent topics | |
| Ebadat et al. | Proper noun semantic clustering using bag-of-vectors | |
| Sengupta | Two models involving Bayesian nonparametric techniques | |
| Zhou et al. | Learning mixed latent tree models | |
| Xu et al. | Improving classification accuracy on uncertain data by considering multiple subclasses | |
| Iwata et al. | Improving classifier performance using data with different taxonomies |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20160713 |